Hacker News new | ask | show | jobs
by ianlevesque 1309 days ago
The structural completions are way more useful than the entire function completions, even in IntelliJ, where autocomplete is already extremely high quality.

The part that I find unsettling when using Copilot is the risk that credentials or secrets embedded in the code, or being edited in (.gitignore'd) config files, are being sent off to Microsoft for AI-munging and possible human review for improvements to the model.
2 comments
patmorgan23 1309 days ago
You shouldn't have any credentials in your git repos anyway. GitHub will already scan your repos and alert you if it thinks there are any credentials in their.
link
ianlevesque 1309 days ago
You've never temporarily put a key into a file while testing? Or accidentally pasted one for a second then deleted it? Can you say the same for your entire team or company?

Since Copilot is constantly making new suggestions, a momentary entry is all it takes.

link
PartiallyTyped 1309 days ago
Copilot doesn't retrain on data generated by you in the moment; so I don't see why this is an issue unless you push the files - with the keys - to github.
link
ianlevesque 1309 days ago
The model is evaluated on the server, using the content of your files.
link
ehutch79 1309 days ago
credentials should never be committed. By the time you're ready to commit code, you should be reading from the environment or a config outside of the codebase, or at least .gitignore'd

Once that key is in your git history, it's in the history. You might be able to edit it, but it's going to be a nightmare to do it.

link
ianlevesque 1309 days ago
I'm not sure why you're referring to committed. The model is being evaluated on the server, with content you haven't yet committed.
link
PartiallyTyped 1309 days ago
> The structural completions are way more useful than the entire function completions, even in IntelliJ, where autocomplete is already extremely high quality.

I needed to run a comparison over a window of a numpy array, and given the sheer size of my data, I needed it to be fast and efficient, which means vectorized operations with minimal python interaction. Copilot figured a solution that is orders of magnitude faster than what I could conjure up in 10 minutes, most of which I'd spent searching for similar solutions in SO.

link